Hedra has launched Model Factory, a post-training and inference platform that uses recursive self-improvement to adapt AI models for robots and other physical machines — and it is available today as a service. Why it matters: every robot task tends to need its own bespoke model, and training thousands of them from scratch is not feasible; Hedra is selling the production line instead of the model.
November 10, 2026
What happened
Hedra Inc. announced the Hedra Model Factory on November 10, according to SiliconANGLE. The platform is designed for robotics, manufacturing, biology and defense workloads. Customers provide a base model and upload proprietary simulation and video data into a GPU sandbox environment powered by Hedra’s inference engine. Hedra pairs its in-house research teams with AI agents that orchestrate training runs, curate data and iterate autonomously to the customer’s specifications.
Founder and CEO Michael Lingelbach, a former reinforcement learning researcher at Stanford and Nvidia, told SiliconANGLE that the goal is to give those teams infrastructure to post-train, deploy and keep improving their own models.
How the self-improvement loop is supposed to work
Recursive self-improvement, or RSI, describes systems that evaluate, train and improve their own architectures in a continuous loop. Hedra says its researchers monitor progress while vision-language models score data and provide live reinforcement learning feedback. Finished models can be deployed on-premises, with a cloud partner, or distilled to run on edge systems for lower-latency work. Once running, they feed data back into the customer’s sandbox, creating what Hedra describes as a continuous optimization flywheel.
Hedra also says RSI can sometimes bypass retraining altogether: in one demonstration, it swapped a robot’s gripper for a new one and restored the task success rate to 100% without additional training, by mapping the new hardware’s reach and grasp metrics to the robot’s wrist camera view.
The validation examples Hedra is citing
The company says it retrained MiniMax’s open-weight H3 video model to improve physical consistency, producing a distilled version that raised its Physics-IQ benchmark score from 30.3 to 34.9. In another run, it distilled the Krea 3 2K image generation model from 30 sampling steps to 14, cutting latency from 22 seconds to 3.6 seconds — a roughly sixfold speed boost that Hedra says took one week.
Those are company-reported results, not independent audits, and physical AI buyers will want to see the same gains on their own hardware and tasks. But they are specific, checkable claims, which is more than most platform launches offer on day one.
Why it matters
Physical AI has a scaling problem that software agents do not: the data is video, sensors and 3D simulation, the compute costs are heavy, and each machine-and-task pairing behaves differently. If Model Factory can reliably turn one base model into many task-specific models — and keep improving them from live data — it addresses the real bottleneck between a impressive robot demo and a fleet that works every day. That is the claim to test, and Hedra says the service is available now for enterprises that want to try.
Frequently asked questions
What is Hedra Model Factory?
A post-training and inference platform, launched November 10, 2026, that uses recursive self-improvement to adapt AI models for robots and other physical systems, offered as a service.
What results has Hedra reported?
Company-reported examples include raising a distilled MiniMax H3 model’s Physics-IQ score from 30.3 to 34.9, and cutting Krea 3 latency from 22 seconds to 3.6 seconds.
Who is it for?
Robotics, manufacturing, biology and defense teams that need bespoke, task-specific models without running compute-intensive training entirely in-house.
Sources: SiliconANGLE